Paper recorded by Signals 4 on 2026-10-19 in cs.CV. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-19 on arXiv · recorded by Signals 4 on 2026-10-20
Category: cs.CV · 计算机视觉 · first seen 2026-10-20
Recent embodied systems increasingly build on multimodal large language models (MLLMs) for their pretrained understanding and reasoning capabilities, but their perception of the physical world remains largely limited to basic modalities such as vision and audio. However, broadening MLLMs to more diverse sensory modalities from the physical environment is challenging: their representations are ofte